ByteDance has been unusually active lately. First, Doubao rolled out a paid professional mode. Then reports emerged that the company was seeking a $20 billion loan overseas. Soon after, ByteDance CEO Liang Rubo sent an all-hands email to employees worldwide. In it, he laid out for the first time a concrete path for the new cycle: “turn computing into intelligence, and use intelligence to improve creativity and experience.”

Liang had previously stressed publicly that over the past few years, ByteDance has been narrowing the scope of its businesses and focusing its energy on AI, with an even tighter focus within AI on improving model capabilities.

Taken together, these moves show that ByteDance has confirmed AI as a strategic direction at the level of first principles. But under that strategy, hope and pressure now stand on both sides of the company.

The hope comes from hard growth in usage. As of June 2026, the Doubao large model was handling an average of 180 trillion tokens a day, more than 1,500 times its early launch level, while the Doubao app had more than 200 million daily active users.

The pressure comes from revenue and costs. Some reports say ByteDance’s AI infrastructure spending in 2026 could reach 200 billion yuan, while others have mentioned annual AI capital expenditure of $70 billion, or about 470 billion yuan.

LatePost also reported that Doubao generates less than 1 million yuan in daily revenue. But based on Volcano Engine’s public API pricing and gross margins, Doubao’s daily inference computing power costs could run into tens of millions of yuan.

Even Seedance, which outsiders have described as ByteDance’s “money printer,” has been played down by the company. Volcano Engine President Tan Dai clarified at a media event last month that online claims of “monthly revenue exceeding 1 billion yuan” were inflated, and that the actual figure was lower.

As ByteDance becomes increasingly committed to AI investment, the outside world keeps returning to one practical question: how can the company earn back the hundreds of billions of yuan it continues to pour in?

ByteDance’s AI Business Turns to Enterprise Customers

Looking back at more than two decades of commercialization in China’s internet industry, outside content consumption categories such as gaming, almost no company has built its core revenue by “charging consumer users directly,” whether in social media, search and recommendation, e-commerce, or later short video.

The same pattern is playing out in AI. In China, products such as ERNIE Bot, Doubao, Yuanbao and Qianwen have yet to produce a scaled revenue model supported by paid memberships. The same is true overseas. AI users globally have generally shown limited willingness to pay; even ChatGPT, the industry leader, has a paid conversion rate of only around 5%.

Traditional traffic platforms ultimately built revenue at the hundreds-of-billions-yuan level through advertising, e-commerce and value-added services. At its core, that remains a business logic in which enterprise customers pay. ByteDance has been one of the beneficiaries of that model. From its founding to becoming one of China’s highest-revenue internet companies, ByteDance’s commercial base has always rested on traffic distribution, with core revenue coming from B2B budgets such as advertising, e-commerce commissions and merchant marketing.

Making money from enterprise customers is also the shared view among AI players.

In AI commercialization, ByteDance has so far put forward two directions: a Doubao Pro subscription, and Volcano Engine’s MaaS offering.

Start with Doubao. On June 24, Doubao launched its professional version, with recurring monthly plans priced at 68 yuan, 200 yuan and 500 yuan, each tied to different usage quotas and feature access.

Doubao grew out of the consumer market. To preserve its user base, ByteDance ultimately avoided an across-the-board paywall and instead created tiers based on model capability. The top-tier features are almost entirely aimed at office use: support for operating a local computer, using a browser, invoking Skills and scheduled tasks, along with a built-in Office productivity suite.

In this setup, Doubao still wears the shell of a consumer product, but the capabilities it offers are already very close to enterprise office needs. In effect, it is filtering B2B users with a willingness to pay for productivity out of consumer traffic ahead of time.

On the enterprise side, the core vehicle for ByteDance’s AI commercialization ambitions is Volcano Engine, whose positioning has continued to shift. ByteDance is steering more resources toward Volcano’s MaaS business and accelerating the rollout of enterprise-grade AI services.

In the product logic of traditional cloud vendors, infrastructure such as computing power, databases, storage and networking sits at the center, while models are an add-on capability within the cloud product matrix.

But ByteDance’s product adjustments and launches over the past two weeks show that large models have become the absolute center of Volcano Engine. Cloud services, agent tools, security systems and nearly all other products are being built around model capabilities and designed to help models land in enterprise scenarios.

ByteDance has also not launched a consumer product in a long time. In May, the X account Mr. Xiaochuan, @xiaochuan8688, posted that according to industry sources, ByteDance had cut 30% of its AI application projects during an internal AI strategy review in April, including Maoxiang, Xinghui and some lines of the overseas AI video tool Dreamina.

The post also said all products outside Doubao had missed expectations; AI video, AI writing and AI education had burned through billions of yuan without producing a breakout product; computing power costs remained difficult to control, with AI inference costs exceeding 8 billion yuan in 2025; and overseas operations were under policy pressure from a potential TikTok divestiture in the U.S., the EU AI Act and India’s ban, narrowing ByteDance’s window for taking AI abroad.

TMTPost previously reported that a ByteDance insider called the information inaccurate. ByteDance has not issued an official response.

In a sense, Volcano Engine has been shaped into a core outlet for ByteDance’s AI capabilities: underneath, it supports self-developed foundation models and a cloud computing power base; above, it connects to real-world product entry points such as Doubao, TRAE, Coze and Jimeng.

Packaging these capabilities into industry solutions that can be deployed directly into the production systems of enterprises, developers and industrial customers is the path ByteDance has found for AI commercialization that best fits its own DNA. After a long turn, ByteDance’s AI business has ultimately returned to the B2B battlefield.

ByteDance’s B2B Test Is a War With a Different Rulebook

The shift to enterprise customers is a consensus the entire industry has reached after doing the math. For ByteDance, the road ahead already has two overseas reference points, one positive and one negative.

One is Anthropic, founded by former core OpenAI team members, and the fastest-growing player in the global large model race over the past two years. In January this year, Anthropic’s annualized revenue was $9 billion. By March it had climbed to $19 billion, and by May it had doubled to $45 billion.

What Anthropic got right was going all in on government and enterprise customers from day one, focusing on high-paying enterprise clients with strong repeat purchase rates. Its B2B revenue has now surpassed OpenAI’s.

On the other side, OpenAI, which used the consumer-facing ChatGPT to open the global market, still has not found a working profit model. In 2025, OpenAI generated $13.07 billion in total revenue, but full-year spending reached $34 billion, leaving an operating loss of about $20.9 billion. The attention and subscription revenue brought by consumer users still cannot cover the huge cost gap created by massive computing power investment.

For ByteDance, these two examples map neatly onto the two cards in its own hand: Doubao Pro and Volcano Engine. But for now, Doubao’s paid tier cannot support the revenue base, and the enterprise fight will not be easy either.

On one hand, Doubao’s 345 million monthly active users look enormous, but the heavy office users and professional creators targeted by the professional version may make up only a small share of the overall user pool. Combined with generally weak willingness to pay for AI tools in China, Doubao Pro is bound to remain a supplement to AI revenue rather than a core pillar.

This sluggish growth had already appeared before the paid policy took effect. Data from third-party platform Aicpb.com showed that Doubao’s monthly active users fell 1.81% month on month in May 2026, a decline of about 6.1 million users.

After the paid version launched, user feedback further confirmed the narrow room for individual subscriptions. Complaints appeared across social platforms, including that “the professional version quota is not enough, and three simple tasks used up 30% of it,” and that “free voice call time has been reduced.”

On the other hand, Volcano Engine faces two giants in Alibaba Cloud and Tencent Cloud. ByteDance’s foundation in enterprise services is not as deep as that of Tencent Cloud or Alibaba Cloud.

The gap is obvious in overall revenue. In 2025, Volcano Engine’s total revenue was about 20 billion yuan, only roughly 60% of Tencent Cloud’s 32 billion to 35 billion yuan as estimated by brokerages, and less than one-seventh of Alibaba Cloud’s 146.6 billion yuan.

Looking only at MaaS, Volcano Engine has indeed delivered rapid growth. IDC data shows it held 49.5% of China’s public cloud MaaS market, close to half the market. But the underlying numbers show that Volcano’s full-category MaaS actual revenue in 2025 was about 1.5 billion yuan. Even if its 2026 revenue target has been raised to 15 billion yuan, Volcano Engine still has a long way to go in establishing its industry position.

A practitioner responsible for enterprise cloud service procurement told Tech Planet, however, that “MaaS alone is not enough.” Volcano still needs to fill gaps in underlying computing power and infrastructure, as well as major-account capabilities, before it can compete head-on with Alibaba Cloud and Tencent Cloud.

Behind the All-Hands Email: ByteDance’s AI War Has Reached the Organization

Liang Rubo’s first all-hands email after taking over as CEO in 2022 was about loosening constraints: stretching bimonthly OKRs into quarterly ones, slowing mature businesses, and emphasizing pragmatism and efficiency.

Four years later, the all-hands email released on June 29 shifted from easing burdens to applying pressure. At its core, ByteDance is rebuilding its organizational logic around its AI strategy. Changes such as refreshing the mission, restructuring management principles and updating 10 leadership principles amount to turning AI strategy from a business-level direction into a hard standard for organizational assessment, with no buffer.

Several core adjustments are highly targeted and map directly to the key pain points in ByteDance’s current AI commercialization push.

The new principles of “doing work with ambition” and “daring to set high goals” are essentially meant to break ByteDance’s past dependence on rapid validation and rapid monetization. AI’s heavy investment and long-cycle nature mean that neither underlying computing power infrastructure nor B2B industry solutions can quickly close the loop the way consumer products can. ByteDance and its employees must accept harder problems and a longer fight.

Making “maintain a sense of crisis and an external perspective” and “go deep into the front line” standalone principles points directly at ByteDance’s weakness in B2B. Facing rivals such as Alibaba Cloud and Tencent Cloud, which have spent more than a decade cultivating industries, Volcano cannot win orders from top enterprise customers on model performance and price competition alone if it does not understand customer needs or have firsthand front-line awareness.

For a large company with more than 100,000 employees, moving from products to services, from traffic thinking to customer thinking, and from rapid iteration to long-cycle delivery requires thorough organizational alignment. Otherwise, strategy can easily get stuck at the execution layer.

This all-hands email therefore became a clear signal that ByteDance has entered the second half of the AI race. As gaps in model capability gradually narrow and commercialization pressure becomes more concrete, technology and organization together will determine the final outcome of the monetization push.